The focus of this paper is the feature extraction from frontal face image. The two main parts are the algorithm of feature points locating (eye locating in particular) and the algorithm of feature extraction. The eye locations are very important landmarks in feature points locating process. They not only indicate and restrict the rough location of other feature points, but also act as significant evidence of normalization of ROI (Region of Interest). Based on the formal researchers’ work, a fast hierarchical-based eye-locating algorithm is proposed here to locate eyes in intensity face image. This new algorithm involves three main steps. First, MER (Minimum Extremal Region) is introduced here to locate eyes roughly. Then, a three-layer-filter is applied to rule out false detections step by step. Finally, an accurate eye-locating method is used to adjust the eye locations precisely. This algorithm has good results in many public face databases. Therefore, it is considered to be an effective method. In the step of feature extraction, this paper pays much attention to the widely-used LBP (Local Binary Pattern) feature. After carefully analysis to LBP and ULBP (Uniform LBP) feature, we propose several improvements on them. In contrary to standard ULBP feature, the Dual-1D LBP feature we proposed brings better results with less feature length. Besides, comparing to single grid on ROI of face image, we use multi grids instead, which can retrieve global, local and trivial information. It can also deduce the linear relativity within feature. Particularly, based on the algorithm research, we implement a face recognition system, which verifies the efficiency of our improvements or proposed methods.
修改评论